void col2blk2(const int M, const int N, const TYPE *A, const int lda, TYPE *V,
const TYPE *alpha)
{
int j;
const int nNb = ATL_DivByNB(N), jb = N - ATL_MulByNB(nNb);
size_t incA = (lda*NB)<<1, incV = (M*NB)<<1;
for (j=nNb; j; j--)
{
col2blk(M, NB, A, lda, V, alpha);
A += incA;
V += incV;
}
if (jb) col2blk(M, jb, A, lda, V, alpha);
}
void Mjoin(PATL,pcol2blkF)
(const int M, const int N, const SCALAR alpha, const TYPE *A, int lda,
const int ldainc, TYPE *V)
/*
* Copies entire MxN matrix to block major format
*/
{
int j, jb;
const int incV = ATL_MulByNB(M);
const enum PACK_UPLO UA = (ldainc == 1) ? PackUpper :
( (lda == -1) ? PackLower : PackGen );
void (*col2blk)(const int M, const int N, const SCALAR alpha, const TYPE *A,
int lda, const int ldainc, TYPE *V);
if (ldainc)
{
if (alpha == ATL_rone) col2blk = Mjoin(PATL,pcol2blk_a1);
else col2blk = Mjoin(PATL,pcol2blk_aX);
for (j=0; j < N; j += NB)
{
jb = N-j;
jb = Mmin(jb, NB);
col2blk(M, jb, alpha, A+MindexP(UA,0,j,lda), Mpld(UA,j,lda), ldainc,V);
V += incV;
}
}
else if (alpha == ATL_rone)
Mjoin(PATL,col2blk2_a1)(M, N, A, lda, V, alpha);
else
Mjoin(PATL,col2blk2_aX)(M, N, A, lda, V, alpha);
}
void Mjoin(pcol2blkF,_blk)
(const int blk, const int M, const int N, const SCALAR alpha, const TYPE *A,
int lda, const int ldainc, TYPE *V)
/*
* Copies entire MxN matrix to block major format
*/
{
int j, jb;
const int incV = blk*(M+M);
const enum PACK_UPLO UA = (ldainc == 1) ? PackUpper :
( (lda == -1) ? PackLower : PackGen );
void (*col2blk)(const int blk, const int M, const int N, const SCALAR alpha,
const TYPE *A, int lda, const int ldainc, TYPE *V);
#ifdef Conj_
if (alpha[1] == ATL_rzero)
{
if (*alpha == ATL_rone) col2blk = Mjoin(Mjoin(PATL,pcol2blkConj_a1),_blk);
else col2blk = Mjoin(Mjoin(PATL,pcol2blkConj_aXi0),_blk);
}
else col2blk = Mjoin(Mjoin(PATL,pcol2blkConj_aX),_blk);
#else
if (alpha[1] == ATL_rzero)
{
if (*alpha == ATL_rone) col2blk = Mjoin(Mjoin(PATL,pcol2blk_a1),_blk);
else col2blk = Mjoin(Mjoin(PATL,pcol2blk_aXi0),_blk);
}
else col2blk = Mjoin(Mjoin(PATL,pcol2blk_aX),_blk);
#endif
for (j=0; j < N; j += blk)
{
jb = N-j;
jb = Mmin(jb, blk);
col2blk(blk, M, jb, alpha, A+MindexP(UA,0,j,lda), Mpld(UA,j,lda),
ldainc, V);
V += incV;
}
}